CoreWeave Deploys Multi-Rack Nvidia Vera Rubin Clusters with Enhanced AI Object Storage Capabilities

CoreWeave has deployed multi-rack Nvidia Vera Rubin clusters on its cloud platform. The setup enables training and inference jobs to run across hundreds of Rubin GPUs, with each rack containing 72 Rubin GPUs.

A single Nvidia Vera Rubin NVL72 rack includes 36 Vera CPUs, Nvidia NVLink 6, Nvidia ConnectX-9 SuperNICs, and Nvidia BlueField-4 DPUs. The company connects multiple racks using Nvidia Spectrum-X Ethernet networking.

CoreWeave's product and engineering executive vice president, Chen Goldberg, stated that the company was the first AI cloud provider to validate and bring up a Vera Rubin NVL72. He also mentioned that with multi-rack Vera Rubin, hundreds of Rubin GPUs can be connected as a single scale-out cluster.

The announcement comes along with two new capabilities in CoreWeave's AI Object Storage: cross-region write acceleration and a new Archive tier providing lower-cost storage with no retrieval or reading fees. The LOTA system provides up to 7 GB/s of throughput per GPU, delivering reads at local NVMe speeds and reducing latency by 8x compared to traditional storage clusters.

Cohere's director of internal infrastructure, Cécile Robert-Michon, said that CoreWeave AI Object Storage gives them a unified dataset footprint across regions with reads cached locally. She noted that this results in nothing waiting on the network.